Let's talk about AI tools for drafting a patent application
Over the past two years, tools like ChatGPT, Gemini and other general-purpose large language models (LLMs) have been marketed as shortcuts for drafting patent applications. For non‑specialists, it is tempting: paste in a description, get a “patent application” back in seconds. The problem is that, in real patent practice, these tools are structurally bad at many of the things that matter most — including sufficiency of disclosure, technical accuracy, legal framing, and portfolio strategy.
Patent offices are increasingly explicit that AI outputs must not be trusted blindly. The USPTO’s 2024 guidance on AI tools states that “relying on the accuracy of an AI tool is unreasonable” and reminds practitioners that every paper filed must be reviewed and verified by a human, with a duty of reasonable inquiry and candour. The EPO’s 2026 Guidelines similarly emphasise that parties are responsible for the content of applications and submissions “regardless of whether a document has been prepared with the assistance of an artificial intelligence (AI) tool.”
In other words: you can’t outsource judgement, responsibility, or quality control to a general-purpose chatbot — and the case law already shows what goes wrong when disclosure is vague, generic, or treated as a black box.
How LLM-Style Drafting Collides with Legal Requirements
1. Generic descriptions and “black box” disclosure
Modern patent law requires more than attractive prose; it demands specific technical teaching that enables a skilled person to carry out the invention and demonstrates that the inventor genuinely possessed it. In both the US and Europe, this translates into written description and sufficiency of disclosure requirements that AI-style generic drafting often fails to meet.
Several recent decisions illustrate the problem - thanks to Finnegan for the leg work here!:
In Ex parte Buhrmann (PTAB, 2020), claims on authenticating electronic transactions were rejected for lack of written description because the specification described only generic “weighted parameters” and high-level outcomes, without a specific algorithm for determining the claimed risk score. The Board held that the disclosure was too generic to show possession of the invention.
In Ex parte Allen (PTAB, 2020), AI was used conceptually to build patient-specific medication lists, but the specification relied on known AI techniques in broad terms, again without concrete implementation detail. The Board found this inadequate to support the claimed method.
On the European side, decisions T 1539/20, T 0606/21 and T 1526/20 refused applications on AI-related inventions for lack of sufficiency of disclosure, even where counterpart US patents had been granted. In each case, the Boards criticised descriptions that mentioned automated learning systems or deep neural networks but did not explain, in practical detail, how the system actually achieved the claimed technical effect.
None of these decisions say “this was drafted by ChatGPT”, but they are textbook examples of the kind of generic, high-level language that general-purpose LLMs tend to produce: descriptions of goals and buzzwords, not concrete teaching of how to achieve those goals. When you ask a chatbot to “write a patent about an AI system that evaluates trajectories”, you will typically get exactly the kind of vague, result-focused language that the Boards of Appeal and PTAB are rejecting.
2. Hallucinations and fabricated references
LLMs are trained to produce plausible text, not verified truth. They are known to hallucinate citations, technical details and prior‑art references, especially in specialised domains like patents. The USPTO’s AI guidance explicitly warns that AI tools “sometimes ‘hallucinate’ or output incorrect information” and stresses that all citations, technical information, priority claims and factual assertions must be checked by a human before filing.
If a practitioner blindly files an AI‑drafted application that includes fabricated prior‑art citations, incorrect priority data, or technically impossible embodiments, they risk:
Breaching the duty of candour and good faith (specific to US practice)
Making statements that are later used against the patent (e.g. over‑broad technical promises that cannot be substantiated).
Wasting examination cycles correcting errors that should never have been in the file.
General-purpose LLMs are not tuned for patent corpora, legal accuracy, or office practice; they simply do not know which of their own statements are wrong.
3. Misalignment with real portfolio strategy
Chatbots also have no sense of portfolio context: they cannot see how a draft fits within your existing claims, families, jurisdictions, licensing strategy, or competitive landscape unless a human deliberately curates and feeds that information. Even then, they lack:
An understanding of claim scope trade‑offs (what is too broad, too narrow, or strategically misaligned).
Awareness of prosecution history estoppel, divisional timing, and continuation/parent‑child strategy across jurisdictions.finnegan
Sensitivity to commercial realities: which embodiments actually matter for your products, customers and exit scenarios.
When companies ask a generalist LLM to “draft a patent” in isolation, they typically end up with unprioritised claims, vague embodiments and no clear link to commercial objectives — a weak foundation for a serious portfolio.
There Is No “AI Inventor”: Law Is Still Human-Centric
The most visible AI-and-patent case law to date underlines a simple truth: patents are still fundamentally about human inventors and human responsibility.
The EPO’s J 8/20 (DABUS) decision confirmed that an AI system cannot be named as inventor under the European Patent Convention; the inventor must be a natural person with legal capacity. The USPTO’s inventorship guidance for AI-assisted inventions likewise insists that patents must list natural persons who made a “significant contribution” to the invention, and that AI tools, however powerful, do not replace human inventors.
All filings must be signed by a human, who certifies they have made a reasonable inquiry and that the contents are accurate to the best of their knowledge.
Parties remain fully responsible for compliance with EPC, PCT and national law, regardless of whether an AI tool assisted in drafting.
Taken together, these positions make one thing clear: AI may help with supporting tasks, but the legal system still expects trained humans to own the thinking, the accountability and the quality.
The Gold Standard: Trained Human Expertise
For now — and likely for a long time — the gold standard in patent drafting remains a trained human: a patent attorney, patent agent or specialised patent consultant with both technical and legal depth.
A competent human patent professional brings capabilities that general-purpose LLMs simply do not have:
Technical comprehension and dialogue: the ability to sit with scientists and engineers, probe assumptions, identify the real inventive core and map it to claim language that stands up to scrutiny.
Legal framing: understanding sufficiency, support, clarity, added matter, eligibility and obviousness in each jurisdiction, and drafting with those constraints in mind from the outset.
Strategic portfolio thinking: designing families, claim sets and filing sequences that support actual business objectives — product protection, licensing, defensive publications, or exit valuation — rather than treating each application as an isolated document.
Ethical and procedural responsibility: knowing when something is too speculative to claim, when prior art is genuinely material, and when a disclosure needs to be strengthened before filing.
General-purpose LLMs can mimic the surface of legal language, but they do not carry this responsibility, nor do they understand the procedural and commercial consequences of the text they generate. That gap is precisely where trained humans earn their keep.
At 8 Bit IP, this is the core of the model: embedded, quasi in‑house IP counsel that combines deep technical fluency (software, electronics, AI, telecoms, physics) with hands‑on experience in research commercialisation, invention harvesting and cross‑border portfolio leadership. The emphasis is on practical, commercially aligned protection — not just “getting something on file”.
Where Specialist AI Tools Do Make Sense
None of this means “no AI ever”. It means that general-purpose chatbots are the wrong tool for high‑stakes patent drafting, whereas specialist, domain‑tuned tools controlled by a human expert can add real value.
Modern specialist platforms built for patent practice — including tools from providers like Solve Intelligence and DeepIP — focus on tasks such as:
Structuring claims and descriptions against current EPO Guidelines and case law.
Highlighting consistency issues between claims, embodiments and drawings.
Surfacing relevant prior art or analogous documents from patent corpora to support drafting and examination preparation.
Helping attorneys stay current with guideline changes (for example, recent EPO updates on AI, claim interpretation and products on the market) while maintaining a human‑centric workflow.
These tools are not a replacement for human judgement; they are power‑assist systems for people who already understand the law and technology. Used properly, they help trained professionals work faster and more consistently, without outsourcing responsibility or core reasoning.
At 8 Bit IP, specialist tools like Solve Intelligence are used exactly in this way: as accelerators for analysis and drafting, under the control of an experienced IP strategist, not as autonomous drafters. They help check alignment with EPO guideline changes, surface relevant prior art, and structure complex deep‑tech disclosures — but every application is still built, reviewed and owned by a human.
Ready to secure your innovations with expert guidance? Whether you’re in Ireland, the US, the UK, or beyond, mastering patent drafting techniques will give you the edge you need to win the IP game. Level up your IP strategy and keep your tech ahead of the pack.
Comments